feat: enable v2 training pipeline with controller parity

[garrett4wade]

Description

Bring GatewayTrainController and RolloutControllerV2 to full parity with v1 controllers, enabling the v2 training pipeline for RL training paths.

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 💥 Breaking change
• 📝 Documentation update
• ♻️ Refactoring
• ⚡ Performance improvement
• ✅ Test coverage improvement

Key Changes

V2 Controller Parity

• Route sglang_remote and vllm_remote to RolloutControllerV2 when config._version == "v2"
• Add thread-safe version management, connect_engine guard address support, and clear_batches RTensor storage eviction to GatewayTrainController
• Direct config_perf_tracer calls to individual workers instead of gateway relay
• Pass staleness_manager to WorkflowExecutor in RolloutControllerV2

AsyncRewardWrapper Lifecycle

• Replace weakref finalization + instance counting with atexit shutdown for all shared executors
• Simplify retry logic and executor recreation with compare-and-swap guard
• Reuse AsyncRewardWrapper instances in math agent workflows instead of creating per-call

HTTP Client Unification

• Use create_httpx_client consistently in workflow_context.py
• Add sock_connect/connect timeouts to aiohttp sessions
• Unify HTTP client session usage across inference/training controllers

Example Configs

• Add agent: section (mode: inline, export_style: individual, turn_discount: 1.0) to all example YAML configs
• Switch default workflow from RLVRWorkflow to MathAgent in gsm8k_rl.py
• Add max_tokens to generation config

Cleanup

• Remove obsolete get_custom_reward_fn and VALID_REWARD_FN from areal/reward/__init__.py
• Remove gateway HTTP helper tests superseded by unified client

Risk Areas

• Breaking: get_custom_reward_fn removed from reward public API — callers using this function will need to import reward functions directly
• Force push: Branch history rewritten (rebase + squash onto latest main)

Checklist

• Pre-commit hooks pass (pre-commit run --all-files)
• New tests added (tests/test_async_reward_wrapper.py)
• Branch is up to date with main
• This PR was created by a coding agent via /create-pr

Test Commands

uv run pytest tests/test_async_reward_wrapper.py
uv run pytest tests/experimental/inference_service/test_controller_version.py
uv run pytest tests/test_examples.py

Skipped suites: GPU/distributed tests (tests/grpo/, tests/torchrun/) — require multi-GPU hardware not available locally.